Windows Report says the toggle lives on Edge's AI settings page, and Microsoft doesn't identify the features it powers. I found no Microsoft announcement or support page that confirms this Canary setting. Treat it as a reported test, not a shipping feature, and treat the interface details as Windows Report's observations.
Where to find it and what it says
Windows Report found the toggle under Settings > Copilot and AI, near the bottom of the page. You can also open edge://settings/ai directly. Canary is Edge's most experimental channel, so the control may change, move or never reach Stable.
The description says Edge can use AI models running directly on your device to power supported features, and that those features might not work when the setting is off. The setting also links to Microsoft's Prompt API documentation, which explains how websites and extensions can access Edge's built-in language model.
The wording is cautious. It says features "might not work", not that they will stop. It names no features, and it doesn't say whether turning the toggle off removes model files or stops downloads.
What local AI Edge already has
Microsoft's developer documentation shows Edge already has several on-device AI pieces. The toggle's link to the Prompt API suggests those pieces are related, but Microsoft hasn't said so.
- Prompt API: this experimental API lets a website or extension prompt a small language model built into Edge. Microsoft lists generating and analyzing text, and building application logic from user input, as uses. It has been a developer preview in Canary and Dev since version 138.0.3309.2.
- Writing Assistance APIs: Microsoft's documentation says these let sites summarize, write and rewrite text. They use the built-in Phi-4-mini model.
- Aion-1.0-Instruct: from version 150.0.4070, Microsoft documents this prerelease model as an option in Canary and Dev. It describes the model as smaller and faster than Phi-4-mini. It also supports CPU inference on devices with weaker GPUs or none.
- Language Detector and Translator APIs: the Microsoft Edge Blog says these identify a text's language and translate between language pairs. They are available in Edge 148 and use on-device, task-specific models.
- On-device speech recognition: the same blog post says the latest Canary and Dev channels include a task-specific model that processes speech locally.
Windows Report says Microsoft hasn't confirmed whether the new toggle controls these APIs or other browser features. Any of these could fall under the switch, but that is unconfirmed.
Hardware and storage limits for the Prompt API
Local models take real resources. For the Prompt API developer preview with Phi-4-mini, Microsoft lists these requirements:
- Windows 10 or 11, or macOS 13.3 or later.
- At least 20 GB free on the volume holding your Edge profile. If free space drops below 10 GB, Microsoft says the model is deleted so other browser features have room.
- 5.5 GB or more of GPU VRAM.
- An unlimited data plan or unmetered connection. The model isn't downloaded over a metered connection.
These limits belong to the documented Prompt API preview. They don't necessarily apply to every on-device feature in Edge or to the new toggle. Aion's CPU-inference support means the GPU requirement shouldn't be applied to it.
Microsoft also says the model is downloaded the first time a site calls an API that needs it. Once downloaded, it is shared across sites. The documentation points to edge://on-device-internals for checking your device performance class and, on Aion builds, which model is active.
How it compares with Chrome
Chrome has a similar setting under chrome://settings/ai. Google's help page is more specific than Edge's reported text. It lists writing help, scam warnings, page summaries and tab organization as possible uses. It also says some AI features don't depend on these models and may keep working when the setting is off.
Google documents what turning the setting off does. Chrome deletes the on-device generative model files and won't download them again while the setting is off. Google also says the setting is per user and per device, and that enterprise use is subject to Chrome Enterprise Policy.
None of that is confirmed for Edge. Edge's reported text says only that features "might not work." Don't assume Edge deletes models, blocks downloads or applies per user.
Notes for administrators
Edge already has a documented enterprise control, the BuiltInAIAPIsEnabled policy. It governs whether pages can use the built-in AI APIs: LanguageModel, Summarization, Writer and Rewriter. It is supported on Windows and macOS from version 138 and applies per profile. It does not apply to a profile signed in with a Microsoft account. If you don't configure it, the APIs stay allowed. Disabling it blocks them, and they return an error when used.
Microsoft doesn't say that policy is the same mechanism as the new toggle. It also doesn't say the policy covers every local model or browser feature. So:
- Don't treat the toggle as a "disable all AI" switch. Cloud-based Copilot features are a separate matter, and Microsoft hasn't said the toggle affects them.
- Don't assume the policy and the toggle do the same thing. Test in a Canary or Dev ring before writing any compliance statement.
- If your goal is limiting web-page access to local models, the documented policy is the control to look at today.
What to watch
Three things would make the setting more useful:
- A list of the features and APIs the toggle covers.
- A statement on whether switching it off deletes stored models and stops new downloads, as Chrome documents.
- Clarity on how it interacts with the enterprise policy.
If you're running Canary and see the switch, you can try it and see what changes. Check edge://on-device-internals before and after to see what the browser reports. Report back, because Microsoft's own description doesn't answer this yet.
The change is modest. It does suggest Microsoft sees a need for a user-facing control as local AI spreads through the browser. Whether the control is substantive or cosmetic depends on what Microsoft documents next.